Crowd Density And Abnormal Behavior Analysis Of Video

In recent years, Massive crowed behavior analysis attracts more and more attention in the field of video surveillance, as the horrible violence event grows. Despite the substantial progress being made in object detection and tracking in recent years, crowed density estimation and behavior analysis remains a non-trivial problem. As a result, the global feature approach is an effective method to solve the problem. Our group has done lots of work about abnormal behavior detection based on optical-flow technology, which has made some progress. In order to improve the accuracy of density estimation and crowed anomaly detection, the paper make much research on lots of features and regression models and proposes a novel approach based on region. Also, the paper applied social force model to video analysis and implemented crowed anomaly detection based on optical-flow and social force model. The main work and contributions are as follows:(1) Research on feature selection and regression models. The segment-based feature, edge-based feature and texture feature such as GLCM and LBP are effective features on the density estimation. The experiment compared the performance of different features under different crowdedness levels and evaluate feature fusion capability. At the meanwhile, the experiment also compared the performance of different regression models under different crowdedness levels. These different regression models are linear regression, PLSA regression, gaussian process regression and random forest regression. The experimental results show that different features and regression models should be selected to achieve best accuracy under different crowdedness levels.(2) Crowd density estimation based on region. This paper proposed a novel approach by computing crowed density estimation using different features and regression models, and deeply mining the best feature and regression model under different crowdedness levels. The method overcame the limitation of single regression approach and the experiment shows that the accuracy improved5.8%.(3) Crowd anomaly detection based on optical-flow and social force model. This paper estimates motion using optical-flow and computes interactive force between particles by applying social force model. The method doesn’t depend on pedestrian tracking based on individual target and is suitable for low density crowd scene and high density crowd scene.The algorithm proposed in this paper has been realized and the modules of crowd density estimation and crowd anomaly detection have been playing a vital part in the project of "The Technology and Working Products of Event Detection with Vision Cloud".